Search

Common Voice és un projecte que convida a la gent a donar la seva veu en la seva llengua materna amb el propòsit de reunir corpus de veus. Aquests corpus, disponibles en més de noranta idiomes, estan formats per una gran quantitat d’àudios curts, en els quals els usuaris es graven llegint una frase, i es poden utilitzar per millorar els sistemes de reconeixement i síntesi de parla. Precisament en català la majoria d’aquests sistemes actualment necessiten millores en els seus models i la base de dades de Common Voice pot arribar a ser de gran utilitat. Un dels seus avantatges és l'accessibilitat per a tothom, tant per la gent que vol desenvolupar sistemes de reconeixement com per la que vol proporcionar la seva veu. Per aquest motiu, aquest projecte utilitza el corpus català amb la intenció de validar-lo, intentant generar una segmentació fonètica dels àudios i posteriorment analitzant la precisió de les etiquetes, amb l’objectiu de poder contribuir en el desenvolupament de les tecnologies de la parla, a més d’adquirir coneixements sobre el funcionament de gestió de dades i entrenament de models amb xarxes neuronals. Inicialment, es pretenia processar la base de dades completa, formada per 25 GB de dades i més de 1000 hores gravades, però això requereix una gran quantitat de temps i de recursos computacionals, així que s’ha optat per seleccionar un conjunt reduït de dades i treballar a menor escala. El procediment ha sigut realitzar un alineament forçat amb els àudios i les seves transcripcions amb els programes Festival i Montreal Forced Aligner. Per aquest últim s’ha necessitat entrenar un model acústic, utilitzant el mateix corpus de Common Voice però amb diferents arxius. Finalment s’ha analitzat la precisió temporal de les etiquetes resultants, en les que s’ha vist que no eren del tot perfectes. Common Voice es un proyecto que invita a la gente a donar su voz en su lengua materna con el propósito de reunir corpus de voz. Estos corpus, disponibles en más de noventa idiomas, estan formados por una gran cantidad de audios cortos, en los que los usuarios se graban leyendo una frase, y se pueden utilizar para mejorar los sistemes de reconocimiento y síntesi del habla. Precisamente en catalán la mayoría de estos sistemas actualmente necesitan mejoras en sus models y la base de datos de Common Voice puede llegar a ser de gran utilidad. Una de sus ventajas es la accesibilidad para todo el mundo, tanto para la gente que quiere desarrollar sistemas de reconocimiento como para la que quiere proporcionar su voz. Por eso, este proyecto utiliza el corpus catalán con la intención de validarlo, intentando generar una segmentación fonética de los audios y posteriormente analizando la precisión de las etiquetas, con el objetivo de poder contribuir con el desarrollo de las tecnologías del habla, además de adquirir conocimientos sobre el funcionamiento de gestión de datos y entrenamiento de modelos con redes neuronales. Inicialmente, se pretendía procesar la base de datos completa, formada por 25 GB de datos y más de 1000 horas grabadas, pero esto requiere una gran cantidad de tiempo y de recursos computacionales, así que se ha optado por seleccionar un conjunto reducido de datos y trabajar a menor escala. El procedimiento ha sido realizar un alineamiento forzado con los audios y sus transcripciones con los programas Festival y Montreal Forced Aligner. Para este último se ha necesitado entrenar un modelo acústico, usando el mismo corpus de Common Voice pero con diferentes archivos. Finalmente se ha analizado la precisión temporal de las etiquetas resultantes, en las que se ha visto que no eran del todo perfectas. Common Voice is a project that invites people to donate their voice in their mother tongue in order to collect speech corpora. These corpora, available in more than ninety languages, consist of a large quantity of short audio files, in which users record themselves reading a sentence, and can be used to improve speech recognition and synthesis systems. In catalan, most of these systems currently need improvement in their models and the Common Voice dataset may be very helpful. One of its advantages is the accessibility for everyone, both for people who want to develop speech recognition systems and for people who want to provide their voice data. For this reason, this project uses the catalan corpus with the purpose of validating it, generating a phonetic segmentation of the audios and subsequently analyzing the labeling precision, with the aim of being able to contribute in the development of speech technologies, as well as gaining knowledge about data management and model training. The initial idea was processing the whole dataset, consisting of 25 GB of data and more than 1000 hours of recording, but this requires a large amount of time and computing resources, so a later choice was selecting a reduced set of audios and work on a smaller scale. The procedure was performing forced alignment on the audios and their orthographic transcriptions with Festival and Montreal Forced Aligner. For the latter, the training of a new acoustic model was needed, using the same Common Voice corpus but with different files. Finally the accuracy of the labels was analyzed, but the results showed that they were not entirely accurate.

Models based on attention mechanisms have shown unprecedented speech recognition performance. However, they are computationally expensive and unnecessarily complex for keyword spotting, a task targeted to small-footprint devices. This work explores the application of Lambda networks, an alternative framework for capturing long-range interactions without attention, for the keyword spotting task. We propose a novel \textit{ResNet}-based model by swapping the residual blocks by temporal Lambda layers. Furthermore, the proposed architecture is built upon uni-dimensional temporal convolutions that further reduce its complexity. The presented model does not only reach state-of-the-art accuracies on the Google Speech Commands dataset, but it is 85% and 65% lighter than its Transformer-based (KWT) and convolutional (Res15) counterparts while being up to 100 times faster. To the best of our knowledge, this is the first attempt to explore the Lambda framework within the speech domain and therefore, we unravel further research of new interfaces based on this architecture., Comment: speech recognition, keyword spotting, lambda networks

Recent advances in Deep Learning (DL) for speaker recognition have improved the performance but are constrained to the need of labels for the background data, which is difficult in prac- tice. In i-vector based speaker recognition, cosine (unsuper- vised) and PLDA (supervised) are the basic scoring techniques, with a big performance gap between the two. In this thesis we tried to fill this gap without using speaker labels in several ways. We applied Restricted Boltzmann Machine (RBM) vectors for the tasks of speaker clustering and tracking in TV broadcast shows. The experiments on AGORA database show that us- ing this approach we gain a relative improvement of 12% and 11% for speaker clustering and tracking tasks, respectively. We also applied DL techniques in order to increase the discrimina- tive power of i-vectors in speaker verification task, for which we have proposed the use of autoencoder in several ways, i.e., (1) as a pre-training for a Deep Neural Network (DNN), (2) as a near- est neighbor autoencoder for i-vectors, (3) as an average pooled nearest neighbor autoencoder. The experiments on VoxCeleb database show that we gain a relative improvement of 21%, 42% and 53%, using the three systems respectively. Finally we also proposed a self-supervised end-to-end speaker verification system. The architecture is based on a Convolutional Neural Network (CNN), trained as a siamese network with multiple branches. From the results we can see that our system shows comparable performance to a supervised baseline This work was supported by the project PID2019-107579RBI00 / AEI / 10.13039/501100011033

Most state-of-the-art Deep Learning systems for text-independent speaker verification are based on speaker embedding extractors. These architectures are commonly composed of a feature extractor front-end together with a pooling layer to encode variable-length utterances into fixed-length speaker vectors. In this paper we present Double Multi-Head Attention (MHA) pooling, which extends our previous approach based on Self MHA. An additional self attention layer is added to the pooling layer that summarizes the context vectors produced by MHA into a unique speaker representation. This method enhances the pooling mechanism by giving weights to the information captured for each head and it results in creating more discriminative speaker embeddings. We have evaluated our approach with the VoxCeleb2 dataset. Our results show 6.09% and 5.23% relative improvement in terms of EER compared to Self Attention pooling and Self MHA, respectively. According to the obtained results, Double MHA has shown to be an excellent approach to efficiently select the most relevant features captured by the CNN-based front-ends from the speech signal. This work was supported by the Spanish project PID2019-107579RBI00 / AEI / 10.13039/501100011033.

Speech enhancement improves recorded voice utterances to eliminate noise that might be impeding their intelligibility or compromising their quality. Typical speech enhancement systems are based on regression approaches that subtract noise or predict clean signals. Most of them do not operate directly on waveforms. In this work, we propose a generative approach to regenerate corrupted signals into a clean version by using generative adversarial networks on the raw signal. We also explore several variations of the proposed system, obtaining insights into proper architectural choices for an adversarially trained, convolutional autoencoder applied to speech. We conduct both objective and subjective evaluations to assess the performance of the proposed method. The former helps us choose among variations and better tune hyperparameters, while the latter is used in a listening experiment with 42 subjects, confirming the effectiveness of the approach in the real world. We also demonstrate the applicability of the approach for more generalized speech enhancement, where we have to regenerate voices from whispered signals.

In the last years, i-vectors followed by cosine or PLDA scoringtechniques were the state-of-the-art approach in speaker veri-fication. PLDA requires labeled background data, and thereexists a significant performance gap between the two scoringtechniques. In this work, we propose to reduce this gap by us-ing an autoencoder to transform i-vector into a new speaker vec-tor representation, which will be referred to as ae-vector. Theautoencoder will be trained to reconstruct neighbor i-vectors in-stead of the same training i-vectors, as usual. These neighbori-vectors will be selected in an unsupervised manner accordingto the highest cosine scores to the training i-vectors. The evalua-tion is performed on the speaker verification trials of VoxCeleb-1 database. The experiments show that our proposed ae-vectorsgain a relative improvement of 42% in terms of EER comparedto the conventional i-vectors using cosine scoring, which fillsthe performance gap between cosine and PLDA scoring tech-niques by 92%, but without using speaker labels

Most state-of-the-art Deep Learning (DL) approaches for speaker recognition work on a short utterance level. Given the speech signal, these algorithms extract a sequence of speaker embeddings from short segments and those are averaged to obtain an utterance level speaker representation. In this work we propose the use of an attention mechanism to obtain a discriminative speaker embedding given non fixed length speech utterances. Our system is based on a Convolutional Neural Network (CNN) that encodes short-term speaker features from the spectrogram and a self multi-head attention model that maps these representations into a long-term speaker embedding. The attention model that we propose produces multiple alignments from different subsegments of the CNN encoded states over the sequence. Hence this mechanism works as a pooling layer which decides the most discriminative features over the sequence to obtain an utterance level representation. We have tested this approach for the verification task for the VoxCeleb1 dataset. The results show that self multi-head attention outperforms both temporal and statistical pooling methods with a 18\% of relative EER. Obtained results show a 58\% relative improvement in EER compared to i-vector+PLDA., 4+1 pages. 4 Figures. Accepted for Interspeech 2009

Alzheimer’s disease (AD) is the most prevalent form of degenerative dementia; it has a high socio-economic impact in Westerncountries. The purpose of our project is to contribute to earlier diagnosis of AD and allow better estimates of its severity by usingautomatic analysis performed through new biomarkers extracted through non-invasive intelligent methods. The method selectedis based on speech biomarkers derived from the analysis of spontaneous speech (SS). Thus the main goal of the present work isfeature search in SS, aiming at pre-clinical evaluation whose results can be used to select appropriate tests for AD diagnosis. Thefeature set employed in our earlier work offered some hopeful conclusions but failed to capture the nonlinear dynamics of speechthat are present in the speech waveforms. The extra information provided by the nonlinear features could be especially useful whentraining data is limited. In this work, the fractal dimension (FD) of the observed time series is combined with linear parameters inthe feature vector in order to enhance the performance of the original system while controlling the computational cost.© 2014 Elsevier Ltd. All rights reserved.

Speech Recognition (reconocimiento de voz) es la tarea que pretende indentificar palabras habladas y convertirlas a texto. Este trabajo de fin de grado se centra en utilizar técnicas de deep learning para construir un sistema de Speech Recognition entrenándolo end-to-end. Como paso preliminar, hacemos un resumen de los métodos más relevantes llevados a cabo los últimos años. A continuación estudiamos uno de los trabajos más recientes en este área que propone un modelo sequence to sequence con atención entrenado end-to-end. Después, reproducimos satisfactoriamente el modelo y lo avaluamos con la base de datos TIMIT. Analizamos los parecidos y diferencias entre la implementación propuesta y el trabajo teórico original. Y finalmente, experimentamos y contrastamos el modelo utilizando diferentes parámetros (e.g. numero de neuronas por capa, la tasa de aprendizaje -learning rate y los batch sizes) y reducimos el Phoneme Error Rate cerca del 12% relativo. Speech recognition is the task aiming to identify words in spoken language and convert them into text. This bachelor's thesis focuses on using deep learning techniques to build an end-to-end Speech Recognition system. As a preliminary step, we overview the most relevant methods carried out over the last several years. Then, we study one of the latest proposals for this end-to-end approach that uses a sequence to sequence model with attention-based mechanisms. Next, we successfully reproduce the model and test it over the TIMIT database. We analyze the similarities and differences between the current implementation proposal and the original theoretical work. And finally, we experiment and contrast using different parameters (e.g. number of layer units, learning rates and batch sizes) and reduce the Phoneme Error Rate in almost 12% relative. Speech Recognition (reconeixement de veu) és la tasca que pretén indentificar paraules del llenguatge parlat i convertir-les a text. Aquest treball de fi de grau es centra en utilitzar tècniques de deep learning per construir un sistema d'Speech Recognition entrenant-lo end-to-end. Com a pas preliminar, fem un resum dels mètodes més rellevants duts a terme els últims anys. A continuació, estudiem un dels treballs més recents en aquesta àrea que proposa un model sequence to sequence amb l?atenció entrenat end-to-end. Després, reproduim satisfactòiament el model i l'avaluem amb la base de dades TIMIT. Analitzem les semblances i diferències entre l'implementació proposada i el treball teòric original. I finalment, experimentem i contrastem el model utilitzant diferents paràmetres (e.g. nombre de neurones per capa, la taxa d'aprenentatge -learning rate- i els batch sizes) i reduim el Phoneme Error Rate gairebé un 12% relatiu.

Restricted Boltzmann Machines (RBMs) have shown success in both the front-end and backend of speaker verification systems. In this paper, we propose applying RBMs to the front-end for the tasks of speaker clustering and speaker tracking in TV broadcast shows. RBMs are trained to transform utterances into a vector based representation. Because of the lack of data for a test speaker, we propose RBM adaptation to a global model. First, the global model&mdash, which is referred to as universal RBM&mdash, is trained with all the available background data. Then an adapted RBM model is trained with the data of each test speaker. The visible to hidden weight matrices of the adapted models are concatenated along with the bias vectors and are whitened to generate the vector representation of speakers. These vectors, referred to as RBM vectors, were shown to preserve speaker-specific information and are used in the tasks of speaker clustering and speaker tracking. The evaluation was performed on the audio recordings of Catalan TV Broadcast shows. The experimental results show that our proposed speaker clustering system gained up to 12% relative improvement, in terms of Equal Impurity (EI), over the baseline system. On the other hand, in the task of speaker tracking, our system has a relative improvement of 11% and 7% compared to the baseline system using cosine and Probabilistic Linear Discriminant Analysis (PLDA) scoring, respectively.

This work presents a study on the suitability of prosodic andacoustic features, with a special focus on i-vectors, in expressivespeech analysis and synthesis. For each utterance of two dif-ferent databases, a laboratory recorded emotional acted speech,and an audiobook, several prosodic and acoustic features are ex-tracted. Among them, i-vectors are built not only on the MFCCbase, but also on F0, power and syllable durations. Then, un-supervised clustering is performed using different feature com-binations. The resulting clusters are evaluated calculating clus-ter entropy for labeled portions of the databases. Additionally,synthetic voices are trained, applying speaker adaptive training,from the clusters built from the audiobook. The voices are eval-uated in a perceptual test where the participants have to edit anaudiobook paragraph using the synthetic voices.The objective results suggest that i-vectors are very use-ful for the audiobook, where different speakers (book charac-ters) are imitated. On the other hand, for the laboratory record-ings, traditional prosodic features outperform i-vectors. Also,a closer analysis of the created clusters suggest that differentspeakers use different prosodic and acoustic means to conveyemotions. The perceptual results suggest that the proposed i-vector based feature combinations can be used for audiobookclustering and voice training.

The synthesis of talking heads has been a flourishing research area over the last few years. Since human beings have an uncanny ability to read people's faces, most related applications (e.g., advertising, video-teleconferencing) require absolutely realistic photometric and behavioral synthesis of faces. This paper proposes a person-specific facial synthesis framework that allows high realism and includes a novel way to control visual emphasis (e.g., level of exaggeration of visible articulatory movements of the vocal tract). There are three main contributions: a geodesic interpolation with visual unit selection, a parameterization of visual emphasis, and the design of minimum size corpora. Perceptual tests with human subjects reveal high realism properties, achieving similar perceptual scores as real samples. Furthermore, the visual emphasis level and two communication styles show a statistical interaction relationship.

Deep Learning has been applied successfully to speech processing. In this paper we propose an architecture for speech synthesis using multiple speakers. Some hidden layers are shared by all the speakers, while there is a specific output layer for each speaker. Objective and perceptual experiments prove that this scheme produces much better results in comparison with sin- gle speaker model. Moreover, we also tackle the problem of speaker interpolation by adding a new output layer (a-layer) on top of the multi-output branches. An identifying code is injected into the layer together with acoustic features of many speakers. Experiments show that the a-layer can effectively learn to interpolate the acoustic features between speakers.

Automatic Speech Recognition (ASR) is becoming increasingly ubiquitous, especially in the mobile segment. Fast and accurate ASR comes at a high energy cost which is not affordable for the tiny power budget of mobile devices. Hardware acceleration can reduce power consumption of ASR systems, while delivering high-performance. In this paper, we present an accelerator for large-vocabulary, speaker-independent, continuous speech recognition. It focuses on the Viterbi search algorithm, that represents the main bottleneck in an ASR system. The proposed design includes innovative techniques to improve the memory subsystem, since memory is identified as the main bottleneck for performance and power in the design of these accelerators. We propose a prefetching scheme tailored to the needs of an ASR system that hides main memory latency for a large fraction of the memory accesses with a negligible impact on area. In addition, we introduce a novel bandwidth saving technique that removes 20% of the off-chip memory accesses issued during the Viterbi search. The proposed design outperforms software implementations running on the CPU by orders of magnitude and achieves 1.7x speedup over a highly optimized CUDA implementation running on a high-end Geforce GTX 980 GPU, while reducing by two orders of magnitude (287x) the energy required to convert the speech into text.

This work aims at creating expressive voices from audiobooks using semantic selection. First, for each utterance of the audiobook an acoustic feature vector is extracted, including iVectors built on MFCC and on F0 basis. Then, the transcription is projected into a semantic vector space. A seed utterance is projected to the semantic vector space and the N nearest neighbors are selected. The selection is then filtered by selecting only acoustically similar data. The proposed technique can be used to train emotional voices by using emotional keywords or phrases as seeds, obtaining training data semantically similar to the seed. It can also be used to read larger texts in an expressive manner, creating specific voices for each sentence. That later application is compared to a DNN predictor, which predicts acoustic features from semantic features. The selected data is used to adapt statistical speech synthesis models. The performance of the technique is analyzed objectively and in a perceptive experiment. In the first part of the experiment, subjects clearly show preference for particular expressive voices to synthesize semantically expressive utterances. In the second part, the proposed method is shown to achieve similar or better performance than the DNN based prediction. Copyright © 2016 ISCA.

Background: Speech segmentation is one of the initial and mandatory phases of language learning. Although some people with aphasia have shown a preserved ability to learn novel words, their speech segmentation abilities have not been explored. Aims: We examined the ability of individuals with chronic aphasia to segment words from running speech via statistical learning. We also explored the relationships between speech segmentation and aphasia severity, and short-term memory capacity. We further examined the role of lesion location in speech segmentation and short-term memory performance. Methods & Procedures: The experimental task was first validated with a group of young adults (n=120). Participants with chronic aphasia (n=14) were exposed to an artificial language and were evaluated in their ability to segment words using a speech segmentation test. Their performance was contrasted against chance level and compared to that of a group of elderly matched controls (n=14) using group and case-by-case analyses. Outcomes & Results: As a group, participants with aphasia were significantly above chance level in their ability to segment words from the novel language and did not significantly differ from the group of elderly controls. Speech segmentation ability in the aphasic participants was not associated with aphasia severity although it significantly correlated with word pointing span, a measure of verbal short-term memory. Case-by-case analyses identified four individuals with aphasia who performed above chance level on the speech segmentation task, all with predominantly posterior lesions and mild fluent aphasia. Their short-term memory capacity was also better preserved than in the rest of the group. Conclusions: Our findings indicate that speech segmentation via statistical learning can remain functional in people with chronic aphasia and suggest that this initial language learning mechanism is associated with the functionality of the verbal short-term memory system and the integrity of the left inferior frontal region.

This paper presents the perceptual experiments that were carried out in order to validate the methodology of transforming expressive speech styles using voice quality (VoQ) parameters modelling, along with the well-known prosody (F0, duration, and energy), from a neutral style into a number of expressive ones. The main goal was to validate the usefulness of VoQ in the enhancement of expressive synthetic speech in terms of speech quality and style identification. A harmonic plus noise model (HNM) was used to modify VoQ and prosodic parameters that were extracted from an expressive speech corpus. Perception test results indicated the improvement of obtained expressive speech styles using VoQ modelling along with prosodic characteristics.

In this paper we present experimental results comparing on-line drawings for control population (left and right hand) as well as Alzheimer disease patients. The drawings have been acquired by means of a digitizing tablet, which acquires time information angles and pressures. Experimental measures based on pressure and in-air movements appear to be significantly different for both groups, even when control population performs the tasks with the non-dominant hand.

Acoustic source localization and sound recognition are common acoustic scene analysis tasks that are usually considered separately. In this paper, a new source localization technique is proposed that works jointly with an acoustic event detection system. Given the identities and the end-points of simultaneous sounds, the proposed technique uses the statistical models of those sounds to compute a likelihood score for each model and for each signal at the output of a set of null-steering beamformers per microphone array. Those scores are subsequently combined to find the MAP-optimal event source positions in the room. Experimental work is reported for a scenario consisting of meeting-room acoustic events, either isolated or overlapped with speech. From the localization results, which are compared with those from the SRP-PHAT technique, it seems that the proposed model-based approach can be an alternative to current techniques for event-based localization. © 2014 IEEE.

Alzheimer's disease (AD) is the most prevalent form of progressive degenerative dementia; it has a high socio-economic impact in Western countries. The purpose of our project is to contribute to earlier diagnosis of AD and better estimates of its severity by using automatic analysis performed through new biomarkers extracted from non-invasive intelligent methods. The methods selected in this case are speech biomarkers oriented to Spontaneous Speech. Thus the main goal of the present work is feature search in Spontaneous Speech oriented to pre-clinical evaluation for the definition of test for AD diagnosis. Nowadays our feature set offers some hopeful conclusions but fails to capture the nonlinear dynamics of speech that are present in the speech waveforms. The extra information provided by the nonlinear features could be especially useful when training data is scarce. In this work, the Fractal Dimension (FD) of the observed time series is combined with lineal parameters in the feature vector in order to enhance the performance of the original system.

Speech signals in real scenario ambient are usually mixed with some other signals, such as noise. This may interfere with posterior signal processing applied to the signals. In this work, a new technique of data denoising is presented using multivariate Empirical Mode Decomposition. Different SNR ratios are tested in order to study the evolution of the improvement of the recovered data. An improvement of the analyzed data is obtained with all the SNR levels tested.

The automatic analysis of speech to detect affective states may improve the way users interact with electronic devices. However, the analysis only at the acoustic level could be not enough to determine the emotion of a user in a realistic scenario. In this paper we analyzed the spontaneous speech recordings of the FAU Aibo Corpus at the acoustic and linguistic levels to extract two sets of features. The acoustic set was reduced by a greedy procedure selecting the most relevant features to optimize the learning stage. We compared two versions of this greedy selection algorithm by performing the search of the relevant features forwards and backwards. We experimented with three classification approaches: Naïve-Bayes, a support vector machine and a logistic model tree, and two fusion schemes: decision-level fusion, merging the hard-decisions of the acoustic and linguistic classifiers by means of a decision tree; and feature-level fusion, concatenating both sets of features before the learning stage. Despite the low performance achieved by the linguistic data, a dramatic improvement was achieved after its combination with the acoustic information, improving the results achieved by this second modality on its own. The results achieved by the classifiers using the parameters merged at feature level outperformed the classification results of the decision-level fusion scheme, despite the simplicity of the scheme. Moreover, the extremely reduced set of acoustic features obtained by the greedy forward search selection algorithm improved the results provided by the full set.

Treball realitzat en el marc d'un programa de mobilitat amb la University of Ljubljana. Faculty of Electrical Engineering [ANGLÈS] The two main objectives of this project are to analyse the efficiency of several techniques widely used among the field of emotion recognition through spoken audio signals, and, secondly, obtain empirical data that proves that it is actually plausible to do so with a more than acceptable performance rate. [CASTELLÀ] Los dos objetivos principales de este PFC son, en primer lugar, describir meticulosamente diversas técnicas utilizadas hoy en día por los desarrolladores del sector multi-disciplinario del reconocimiento de emociones a través del habla. En segundo lugar, y escogiendo nuestra manera de encarar el campo, obtener un amplio volumen de datos que confirmen que es actualmente viable conseguir resultados más que satisfactorios. [CATALÀ] Els dos objectius principals d'aquest PFC són, en primer lloc, descriure meticulosament diverses tècniques emprades avui en dia dins el sector de reconeixement d'emocions a través de senyals d'audio, i , en segon lloc, proveïr un important volum de dades que demostrin que és totalment plausible aconseguir un més que acceptable funcionament

Wavelet derived codification techniques are widespread used in image codifiers. The wavelet based compression methods are adequate for representing transients. In this paper we explore the use of the discrete wavelet transform analysis of biological signals in order to improve the data compression capability of data coders. The wavelet analysis provides a subband decomposition of any signal, and this enables a lossless or a lossy implementation with the same architecture. The signals could range from speech to sounds or music, but the approach is more orientated to other biosignals like medical signals EEG, ECG or discrete series. Experimental results based on wavelet coefficients quantification, show a lossless compression of 2:1 in all kind of signals, with a fidelity, measured using PSNR, from 79dB to 100dB, and lossy results preserving most of the signal waveform, with a compression ratio from 3:1 to 5:1, with a fidelity from 25dB to 35 dB.

In this paper we show how a nonlinear preprocessing of speech signal -with high noise- based on morphological filters improves the performance of robust algorithms for pitch tracking (RAPT). This result happens for a very simple morphological filter. More sophisticated ones could even improve such results. Mathematical morphology is widely used in image processing in where it has found a great amount of applications. Almost all its formulations derived in the two-dimensional framework are easily reformulated to be adapted to onedimensional context.

In this paper we explore the use of non-linear transformations in order to improve the performance of an entropy based voice activity detector (VAD). The idea of using a non-linear transformation comes from some previous work done in speech linear prediction (LPC) field based in source separation techniques, where the score function was added into the classical equations in order to take into account the real distribution of the signal. We explore the possibility of estimating the entropy of frames after calculating its score function, instead of using original frames. We observe that jf signal is clean, estimated entropy is essentially the same; but if signal is noisy transformed frames (with score function) are able to give different entropy if the frame is voiced against unvoiced ones. Experimental results show that this fact permits to detect voice activity under high noise, where simple entropy method fails.

This paper analyzes applications of cumulant analysis in speech processing. The main focus is made on an integral representation for cUl1luJants using integrals involving cyclic products of kernels. These integrals serve as a tool for statistical analysis of c1ifferent statistics appearing in speech processing and speech recognition. As an example. we consider statistics based on the sample correlograrn. A representatiou for the cUlnulants of these statistics using integrals involving cyclic products of kernels is obtained.

This paper presents a thorough study of the impact of morphology derivation on N-gram-based Statistical Machine Translation (SMT) models from English into a morphology-rich language such as Spanish. For this purpose, we define a framework under the assumption that a certain degree of morphology-related information is not only being ignored by current statistical translation models, but also has a negative impact on their estimation due to the data sparseness it causes. Moreover, we describe how this information can be decoupled from the standard bilingual N-gram models and introduced separately by means of a well-defined and better informed feature-based classification task. Results are presented for the European Parliament Plenary Sessions (EPPS) English ¿ Spanish task, showing oracle scores based on to what extent SMT models can benefit from simplifying Spanish morphological surface forms for each Part-Of-Speech category. We show that verb form morphological richness greatly weakens the standard statistical models, and we carry out a posterior morphology classification by defining a simple set of features and applying machine learning techniques. In addition to that, we propose a simple technique to deal with Spanish enclitic pronouns. Both techniques are empirically evaluated and final translation results show improvements over the baseline by just dealing with Spanish morphology. In principle, the study is also valid for translation from English into any other Romance language (Portuguese, Catalan, French, Galician, Italian, etc.). The proposed method can be applied to both monotonic and non-monotonic decoding scenarios, thus revealing the interaction between word-order decoding and the proposed morphology simplification techniques. Overall results achieve statistically significant improvement over baseline performance in this demanding task.

This paper describes a multi-domain text-to-speech (MD-TTS)synthesis strategy for generating speech among different domainsand so increasing the flexibility of high quality TTS systems. Tothat effect, the MD-TTS introduces a flexible TTS architecture thatincludes an automatic domain classification module, which allowsMD-TTS systems to be implemented by different synthesis strate-gies and speech corpus typologies. In this work, the performanceof a corpus-based MD-TTS system is subjectively validated bymeans of several perceptual tests

In this work we present a waveform speech coding system including vector quantization. This system can be seen as a vector version of the scalar ADPCM speech coder. In such system the speech samples are grouped in vectors that are coded by a vector quantizer when its prediction subtraction has been made. This prediction is obtained in the coder by a vector predictor. Due to the non-stationarity of the speech signal, the code must be continuously adapted to the local characteristics of the current input signal of the speech. We propose the use of an adaptive vector predictor that follows the speech statistics variations so the prediction error to be coded presents the minimum dynamic range. On the other side it is wellknown that the prediction error is proportional to the signal energy. To compensate this effect a vector quantizer "gain-shape" model has been proposed, so vectors gain and its shape are separately coded. The obtained empirical results are very promising and exhibit good competitivity with other solutions existing in the literature.

We propose an architecture for speaker recognition. This architecture is independent of the text, robust with the presence of noise, and is based on self organizing maps (SOM) (T. Kohonen, 1984). We compare the performance of this architecture for different parametrizations, different signal to noise ratios, with another method for speaker identification based on the arithmetic harmonic spherity measure on covariance matrices (F. Bimbot and L. Mathan, 1993; J. Hernando et al., 1994)

Spanish is a global language, spoken in a big number of different countries with a big dialectal variability‥ This paper deals with the suitability of using a single multi-dialectal acoustic modeling for all the Spanish variants spoken in Europe and Latin America. This paper deals with the suitability of using a single multi-dialectal acoustic modeling for all the Spanish variants spoken in Europe and Latin America. The objective is two fold. First, it allows to use all the available databases to jointly train and improve the same system. Second, it allows to use a single system for all the Spanish speakers. The paper describes the rule- based phonetic transcription used for each dialectal variant, the selection of the shared and the specific phonemes to be modeled in a multi-dialectal recognition system, and the results of a multi-dialectal system dealing with dialects in and out of the training set.

The article presents a robust representation of speech based on AR modeling of the causal part of the autocorrelation sequence. In noisy speech recognition, this new representation achieves better results than several other related techniques.

Speech dynamic features are routinely used in current speech recognition systems in combination with short-term (static) spectral features. Although many existing speech recognition systems do not weight both kinds of features, it seems convenient to use some weighting in order to increase the recognition accuracy of the system. In the cases that this weighting is performed, it is manually tuned or it consists simply in compensating the variances. The aim of this paper is to propose a method to automatically estimate an optimum state-dependent stream weighting in a continuous density hidden Markov model (CDHMM) recognition system by means of a maximum-likelihood based training algorithm. Unlike other works, it is shown that simple constraints on the new weighting parameters permit to apply the maximum-likelihood criterion to this problem. Experimental results in speaker independent digit recognition show an important increase of recognition accuracy.

We study some speech enhancement algorithms based on the iterative Wiener filtering method due to Lim-Oppenheim (1978), where the AR spectral estimation of the speech is carried out using a second-order analysis. But in our algorithms we consider an AR estimation by means of cumulant analysis. This work extends some preceding papers due to the authors, providing a generalization of third- and fourth-order algorithms by means of adding two parameters in the general expression of Wiener filtering. These parameters allow a better control of their performance. Some results are presented considering AWGN but listening tests give similar performance when other noises (diesel engine) are considered.

A procedure for feature selection in isolated word recognition is discussed. The feature selection is performed in two steps. The first step takes into account the temporal correlation among feature vectors in order to obtain a transformation matrix which projects the initial template of N feature vectors to a new space where they are uncorrelated. This step gives a new template of M feature vectors, where M >

Automatic Speech Recognition (ASR) has experienced a dramatic evolution since pioneer development of Bell Lab’s single-digit recognizer more than 50 years ago. Current ASR systems have taken advantage of the tremendous improvements in AI during the past decade by incorporating Deep Neural Networks into the system and pushing their accuracy to levels comparable to that of humans. This article describes and characterizes a representative ASR system with state-of-the-art accuracy and proposes a hardware platform capable of decoding speech in real-time with a power dissipation close to 1 Watt. The software is based on the so-called hybrid approach with a vocabulary of 200K words and RNN-based language model re-scoring, whereas the hardware consists of a commercially available low-power processor along with two accelerators used for the most compute-intensive tasks. The article shows that high performance can be obtained with very low power, enabling the deployment of these systems in extremely power-constrained environments such as mobile and IoT devices. This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No. 833057), the Spanish State Research Agency under grant TIN2016-75344-R (AEI/FEDER, EU), the ICREA Academia program, and the Spanish MICINN Ministry under grant BES-2017-080605.

Highlights? We design a steganography model of speech-in-speech signals. ? The model, Efficient Wavelet Masking, is based on the masking property of the HAS. ? The core is the efficient sorting of the wavelet coefficients of the secret message. ? EWM and other methods such as LSB, FM, SS, SSA, have been analyzed. ? Statistical transparency is better in EWM than in LSB, FM, SS and SSA. Steganography is the process of hiding information on a host signal. Transparency is referred to the ability to avoid suspicion about the existence of a secret message. The most popular mechanisms for hiding data in audio signals are the Least Significant Bit (LSB) substitution, Frequency Masking (FM), Spread Spectrum (SS), and Shift Spectrum Algorithm (SSA). In this paper, we adapt the Frequency Masking concept using an efficient sorting of the wavelet coefficients of the secret messages and use an indirect LSB substitution for hiding speech signals into speech signals. The experimental results show that the proposed model, the Efficient Wavelet Masking (EWM) scheme, has a hiding capacity significantly higher than the Spread and Shift Spectrum Algorithms and additionally a statistical transparency higher than all of the above mentioned mechanisms. Moreover, the transparency is not dependent of the host signal chosen because the wavelet sorting guarantees the adaptation of the secret message to the host signal.

Cepstral coefficients are widely used in speech recognition. In this paper, we claim that they are not the best way of representing the spectral envelope, at least for some usual speech recognition systems. In fact, cepstrum has several disadvantages: poor physical meaning, need of transformation, and low capacity of adaptation to some recognition systems. In this paper, we propose a new representation that significantly outperforms both mel-cepstrum and LPC-cepstrum techniques in both recognition rate and computational cost. It consists of filtering the frequency sequence of filter-bank energies with an extremely simple filter that equalizes the variance of the cepstral coefficients. Excellent results of the new technique using a continuous observation density HMM recognition system and two very different recognition tasks, connected digits and phone recognition, are presented.

A continuous speech recognition system (called RAMSES) has been built based on the demisyllable as phonetic unit and tools from connected speech recognition. Speech is parameterized by band-pass lifted LPC-cepstra and demisyllables are represented by hidden Markov models (HMM). In this paper, the application of this system to recognize integer numbers from zero to one thousand is described. The paper contains a general overview of the system, an outline of the grammar inference, a description of the HMM training procedure and an assessment on the recognition performance in a speaker independent experiment.

Projecte final de carrera realitzat en col.laboració amb University of Maribor With this diploma work we have attempted to give continuity to the previous work done by other researchers called, Voice Operating Intelligent Wheelchair – VOIC [1]. A development of a wheelchair controlled by voice is presented in this work and is designed for physically disabled people, who cannot control their movements. This work describes basic components of speech recognition and wheelchair control system. Going to the grain, a speech recognizer system is comprised of two distinct blocks, a Feature Extractor and a Recognizer. The present work is targeted at the realization of an adequate Feature Extractor block which uses a standard LPC Cepstrum coder, which translates the incoming speech into a trajectory in the LPC Cepstrum feature space, followed by a Self Organizing Map, which classifies the outcome of the coder in order to produce optimal trajectory representations of words in reduced dimension feature spaces. Experimental results indicate that trajectories on such reduced dimension spaces can provide reliable representations of spoken words. The Recognizer block is left for future researchers. The main contributions of this work have been the research and approach of a new technology for development issues and the realization of applications like a voice recorder and player and a complete Feature Extractor system.

In this article, we present the evaluation results for the task of speaker diarization of broadcast news, which was part of the Albayzin 2010 evaluation campaign of language and speech technologies. The evaluation data consists of a subset of the Catalan broadcast news database recorded from the 3/24 TV channel. The description of five submitted systems from five different research labs is given, marking the common as well as the distinctive system features. The diarization performance is analyzed in the context of the diarization error rate, the number of detected speakers and also the acoustic background conditions. An effort is also made to put the achieved results in relation to the particular system design features.

Treball realitzat en el marc d'un programa de mobilitat amb la University of Ljubljana. Faculty of Electrical Engineering [ANGLÈS] The two main objectives of this project are to analyse the efficiency of several techniques widely used among the field of emotion recognition through spoken audio signals, and, secondly, obtain empirical data that proves that it is actually plausible to do so with a more than acceptable performance rate. [CASTELLÀ] Los dos objetivos principales de este PFC son, en primer lugar, describir meticulosamente diversas técnicas utilizadas hoy en día por los desarrolladores del sector multi-disciplinario del reconocimiento de emociones a través del habla. En segundo lugar, y escogiendo nuestra manera de encarar el campo, obtener un amplio volumen de datos que confirmen que es actualmente viable conseguir resultados más que satisfactorios. [CATALÀ] Els dos objectius principals d'aquest PFC són, en primer lloc, descriure meticulosament diverses tècniques emprades avui en dia dins el sector de reconeixement d'emocions a través de senyals d'audio, i , en segon lloc, proveïr un important volum de dades que demostrin que és totalment plausible aconseguir un més que acceptable funcionament